Archive — CN — technode
DeepSeek releases OCR 2 with new visual encoding architecture, targeting morehuman-like machine vision
Chinese AI startup DeepSeek has launched DeepSeek-OCR 2, an advanced optical character recognition model leveraging a new semantic reasoning architecture for enhanced machine vision. This model reduces computational costs and improves data processing efficiency, showing a 3.73% performance increase in benchmark tests, as competition in AI models intensifies.
Key points
- DeepSeek released DeepSeek-OCR 2 with open-source capability.
- The new model employs DeepEncoder V2 architecture, moving away from rigid visual encoding.
- Semantic reasoning allows for dynamic rearrangement of image components based on context.
- DeepSeek-OCR 2 processes complex documents with fewer visual tokens (256 to 1,120).
- It achieved a 91.09% score on OmniDocBenchv1.5, marking a 3.73% improvement over the previous version.
- This release reflects the intensified competition among Chinese AI developers in foundational models and multimodal AI.
Why it matters
- The release aligns with the 2025 IT trend of improving AI capabilities in document processing and understanding.
- It highlights the ongoing evolution in AI technology driven by competition in large language models.
- DeepSeek's advancement contributes to the broader pattern of increasing open-source initiatives in AI, enabling collaborative innovation.
DeepSeek's release of OCR 2 signifies a significant step toward more human-like machine vision in AI, demonstrating enhanced efficiency and competitive pressure in the AI development landscape.